Speculative fraud hypothesis

[Mauro]

I think it would be interesting to sweep across a wider parameter space for this generator: randomly iterate the code’s various threshold parameters, sweep across a wide range of initializing lines, and re-run a given “seed” (initializing line + set of threshold parameters) multiple times to build up a broader representative population of generated texts. Are all these variations equally convincing, or are some more convincing than others? If there are, how does that relate to the algorithm or initializing line?

In my opinion the seed string does not matter much, if at all: what actually creates the text are the rules, not the seed string, whose effect should peter out quickly. Indeed I'd bet (but I did not try that and can't be sure) that starting from a null string will create a text with the same statistical properties as starting from a proper seed string.

One thing which could be interesting to do is to compare the copy&modify algorithm with a version of it without the copy&modify part, but with the same generation rules. The idea is to generate one word at time, always starting from a null string and applying the rules a random number of times (a few times) for each generated words (one needs also a mechanism for creating 'separable' words: at each step generate two words and decide with a certain probability if to add the 'join two words' rule, else keep only the first word). This could help in separating the effect of the copy&modify mechanism from the effect of the rules.